B-roll Smart Arrangement
This capability is offered through partnership. This page describes the capability only and contains no API parameters. For API integration, contact business@ai-mcn.tv; once confirmed, we provide the full API documentation and integration support.
Arranges your own footage library into B-roll candidate tracks along the beats of your script. It takes over the three most time-consuming parts of laying B-roll:
- Which clip: which piece of which clip goes with each passage;
- Where it goes: which second the shot starts on, and which word it cuts on;
- How long: how the rhythm runs, and what to do when there is not enough footage.
The output is a slot table per track and per slot (which clip, source in/out, track in/out). It drops straight into the Gitruck Cloud client or your own timeline for manual refinement.
Where it sits in the pipeline
Rough-cut talking head / narration → script split into beats → candidate footage retrieved per beat → B-roll Smart Arrangement → captions / music / subtitles → export
It consumes candidates that are already on hand: each beat carries a time range and its highlight keywords; each passage carries a few candidate clips with their matched segments. Candidates can come from Video Material Search, your locally indexed footage, or manual picks; image-quality and highlight signals can be added with Material Understanding. Downstream, the result can go to Project File Export for Jianying / Premiere / FCPX projects, or to Timeline Rendering for a finished video.
How it arranges
Candidate ranking: semantics, aesthetics and highlight fused. For each passage, candidates are scored on "does it match", "does it look good" and "is it worth watching". Blurry or shaky candidates are down-weighted, never excluded: when candidates are scarce, a slightly soft shot beats an empty slot, and the decision to drop stays with you. Candidates you have pinned come first, unconditionally.
Rhythm is a range, not a fixed value. Shot length is drawn from a range around the suggested pace instead of being mechanically equal; re-running the same project reproduces the same result.
Sentence-boundary snapping, roughly 70/30. About seven in ten cuts land on the start of a line and three in ten are deliberately offset: all aligned feels mechanical, all offset feels loose. Dense dialogue skips more, sparse dialogue snaps almost every time; it adapts to density.
Keyword anchors: the picture arrives first. Half a second before a highlight keyword is spoken, its shot is already on screen, rather than "the line ends and then the picture shows up".
No reuse, jump cuts allowed. A given piece of footage is used once per pass; taking different sections of the same clip is legitimate progression and is not treated as a repeat.
Edge trimming against flash frames. When a shot's ends sit close to a scene cut, they are pulled in and snapped to whole frames; shots are only ever shortened, never stretched.
When there is not enough: leave it empty rather than repeat. Three fallback levels: relax and rescan, then leave the slot empty, and after consecutive empties the beat is judged exhausted and laying stops for it. Used footage is never recycled to fill gaps. An optional black bed track can sit underneath.
Multiple candidate tracks are alternative takes. Several candidate tracks can be laid at once; the second track differs from the first in both rhythm and selection, so you can toggle between them on the timeline instead of comparing copies.
Quality checks during arrangement
- Structural self-check, always on: cut points, slivers, minimum shot length and holes are checked one by one. When a candidate has no cut-point data, it says so explicitly ("flash-frame risk undeterminable"), neither calling the unknown safe nor calling it a problem.
- Picture-to-speech alignment, optional: for the key line of each beat, one frame is checked against what the line says; mismatches are re-laid, with a hard cap on rounds. When the cap is reached, the result is delivered and the remainder is recorded honestly.
Checks happen before rendering, so the video is rendered only once.
Footage never leaves your machine
Arrangement needs structured signals only: beat time ranges, candidate scores, matched segments, cut points and motion levels. The footage itself, file paths, scene descriptions and script text are never uploaded. The server keeps no copy of the request body, only scale details and the billing basis.
Billing and limits
Metered by arrangement scale (beats × candidate segments × tracks). The cost can be estimated before execution and capped; over the cap, the request is rejected outright without executing or billing. Resending the same request is not billed twice. Pricing is agreed in the partnership plan.
How to get started
| Route | Best for | What you do |
|---|---|---|
| The gtrk command-line toolbox | Teams that want to start right away | Indexing your footage, retrieval, arrangement, track laying and quality checks are all built in; all you need is an API Key. See the CLI overview |
| API integration | Teams with their own editing pipeline or product | Your side generates beats and candidates and calls the arrangement API; we provide the API documentation, a prerequisites checklist and integration support. Contact sales first: business@ai-mcn.tv |